The emergence of the Internet of Things (IoT) will provide pervasive connectivity for billions of devices in various environments including the so-called emerging smart cities. It is now apparent that – with the growth of urban population from about 50% of world total now to 66% of the world total by 2050  – the IoT will become omnipresent in view of the associated explosion in the number of Internet connected devices sensing, collecting, generating and communicating data in such urban environments. The envisioned applications typically involve the collection of massive amounts of data, related to air pollutants, energy consumption, temperature and humidity, which are then aggregated to the cloud for further analysis. At the same time, social media platforms (Twitter, Facebook, Instagram), RSS feeds, and Open Data Portals provide access to petabytes of diverse, high-dimensional data including text, images and video. The availability of such large-scale, heterogeneous data will offer exciting and unique opportunities – it is widely acknowledged that not only has such big data the potential to transform science, technology, business and commerce via a new wave of more robust, adaptive and personalized technology, but also Future Smart City Ecosystems.

However, the realization of the vision of Smart Cities calls for new effective mechanisms to sense, analyze and process large-scale data to derive actionable knowledge. In this PhD thesis research, we argue that this massive corpus of heterogeneous information can be leveraged by an innovative data processing and analytic pipeline with the potential to significantly contribute to the future of smart city and social media applications, providing for accurate, robust, and novel functionalities, such as monitoring, reporting, and predicting important conditions and events. The envisioned pipeline will leverage the recent advance of graph-based deep learning – a new paradigm in machine learning specially designed for non-Euclidean data – to uncover the underlying correlations and structures of big heterogeneous data.